DBERT is the first BERT model for the Moroccan Arabic dialect called “Darija”. It is based on the same architecture as BERT-base, but without the Next Sentence Prediction (NSP) objective. This model was trained on a total of ~3 Million sequences of Darija dialect representing 691MB of text or a total of ~100M tokens.
The model was trained on a dataset issued from three different sources:
- Stories written in Darija scrapped from a dedicated website
- Youtube comments from 40 different Moroccan channels
- Tweets crawled based on a list of Darija keywords.
More details about DarijaBert are available in the dedicated GitHub repository
Loading the model
The model can be loaded directly using the Huggingface library:
from transformers import AutoTokenizer, AutoModel
DBERT_tokenizer = AutoTokenizer.from_pretrained("Kamel/DBERT")
DBERT_Bert_model = AutoModel.from_pretrained("Kamel/DBERT")
Acknowledgments
We gratefully acknowledge Google’s TensorFlow Research Cloud (TRC) program for providing us with free Cloud TPUs.